Details | Last modification | View Log | RSS feed
Rev | Author | Line No. | Line |
---|---|---|---|
1805 | yogev_ezra | 1 | ; Exports only one function: |
2 | ; void __stdcall lzma_decompress( |
||
3 | ; const void* source, |
||
4 | ; void* destination, |
||
5 | ; unsigned dest_length); |
||
6 | |||
7 | format COFF |
||
8 | |||
9 | section '.text' code |
||
10 | |||
11 | pb equ 2 ; pos state bits |
||
12 | lp equ 0 ; literal pos state bits |
||
13 | lc equ 3 ; literal context bits |
||
14 | posStateMask equ ((1 shl pb)-1) |
||
15 | literalPosMask equ ((1 shl lp)-1) |
||
16 | |||
17 | kNumPosBitsMax = 4 |
||
18 | kNumPosStatesMax = (1 shl kNumPosBitsMax) |
||
19 | |||
20 | kLenNumLowBits = 3 |
||
21 | kLenNumLowSymbols = (1 shl kLenNumLowBits) |
||
22 | kLenNumMidBits = 3 |
||
23 | kLenNumMidSymbols = (1 shl kLenNumMidBits) |
||
24 | kLenNumHighBits = 8 |
||
25 | kLenNumHighSymbols = (1 shl kLenNumHighBits) |
||
26 | |||
27 | LenChoice = 0 |
||
28 | LenChoice2 = 1 |
||
29 | LenLow = 2 |
||
30 | LenMid = (LenLow + (kNumPosStatesMax shl kLenNumLowBits)) |
||
31 | LenHigh = (LenMid + (kNumPosStatesMax shl kLenNumMidBits)) |
||
32 | kNumLenProbs = (LenHigh + kLenNumHighSymbols) |
||
33 | |||
34 | kNumStates = 12 |
||
35 | kNumLitStates = 7 |
||
36 | kStartPosModelIndex = 4 |
||
37 | kEndPosModelIndex = 14 |
||
38 | kNumFullDistances = (1 shl (kEndPosModelIndex/2)) |
||
39 | kNumPosSlotBits = 6 |
||
40 | kNumLenToPosStates = 4 |
||
41 | kNumAlignBits = 4 |
||
42 | kAlignTableSize = (1 shl kNumAlignBits) |
||
43 | kMatchMinLen = 2 |
||
44 | |||
45 | IsMatch = 0 |
||
46 | IsRep = 0xC0 ; (IsMatch + (kNumStates shl kNumPosBitsMax)) |
||
47 | IsRepG0 = 0xCC ; (IsRep + kNumStates) |
||
48 | IsRepG1 = 0xD8 ; (IsRepG0 + kNumStates) |
||
49 | IsRepG2 = 0xE4 ; (IsRepG1 + kNumStates) |
||
50 | IsRep0Long = 0xF0 ; (IsRepG2 + kNumStates) |
||
51 | PosSlot = 0x1B0 ; (IsRep0Long + (kNumStates shl kNumPosBitsMax)) |
||
52 | SpecPos = 0x2B0 ; (PosSlot + (kNumLenToPosStates shl kNumPosSlotBits)) |
||
53 | Align_ = 0x322 ; (SpecPos + kNumFullDistances - kEndPosModelIndex) |
||
54 | Lencoder = 0x332 ; (Align_ + kAlignTableSize) |
||
55 | RepLencoder = 0x534 ; (Lencoder + kNumLenProbs) |
||
56 | Literal = 0x736 ; (RepLencoder + kNumLenProbs) |
||
57 | |||
58 | LZMA_BASE_SIZE = 1846 ; must be ==Literal |
||
59 | LZMA_LIT_SIZE = 768 |
||
60 | |||
61 | kNumTopBits = 24 |
||
62 | kTopValue = (1 shl kNumTopBits) |
||
63 | |||
64 | kNumBitModelTotalBits = 11 |
||
65 | kBitModelTotal = (1 shl kNumBitModelTotalBits) |
||
66 | kNumMoveBits = 5 |
||
67 | |||
68 | RangeDecoderBitDecode: |
||
69 | ; in: eax->prob |
||
70 | ; out: CF=bit |
||
71 | push edx |
||
72 | mov edx, [range] |
||
73 | shr edx, kNumBitModelTotalBits |
||
74 | imul edx, [eax] |
||
75 | cmp [code_], edx |
||
76 | jae .ae |
||
77 | mov [range], edx |
||
78 | mov edx, kBitModelTotal |
||
79 | sub edx, [eax] |
||
80 | shr edx, kNumMoveBits |
||
81 | add [eax], edx |
||
82 | .n: |
||
83 | pushfd |
||
84 | call update_decoder |
||
85 | popfd |
||
86 | pop edx |
||
87 | ret |
||
88 | .ae: |
||
89 | sub [range], edx |
||
90 | sub [code_], edx |
||
91 | mov edx, [eax] |
||
92 | shr edx, kNumMoveBits |
||
93 | sub [eax], edx |
||
94 | stc |
||
95 | jmp .n |
||
96 | |||
97 | update_decoder: |
||
98 | cmp byte [range+3], 0 ;cmp dword [range], kTopValue |
||
99 | jnz @f ;jae @f |
||
100 | shl dword [range], 8 |
||
101 | shl dword [code_], 8 |
||
102 | push eax |
||
103 | mov eax, [inptr] |
||
104 | mov al, [eax] |
||
105 | inc dword [inptr] |
||
106 | mov byte [code_], al |
||
107 | pop eax |
||
108 | @@: ret |
||
109 | |||
110 | LzmaLenDecode: |
||
111 | ; in: eax->prob, edx=posState |
||
112 | ; out: ecx=len |
||
113 | |||
114 | ; LenChoice==0 |
||
115 | ; add eax, LenChoice*4 |
||
116 | call RangeDecoderBitDecode |
||
117 | jnc .0 |
||
118 | add eax, (LenChoice2-LenChoice)*4 |
||
119 | call RangeDecoderBitDecode |
||
120 | jc @f |
||
121 | mov cl, kLenNumMidBits |
||
122 | shl edx, cl |
||
123 | lea eax, [eax + (LenMid-LenChoice2)*4 + edx*4] |
||
124 | call RangeDecoderBitTreeDecode |
||
125 | add ecx, kLenNumLowSymbols |
||
126 | ret |
||
127 | @@: |
||
128 | add eax, (LenHigh-LenChoice2)*4 |
||
129 | mov cl, kLenNumHighBits |
||
130 | call RangeDecoderBitTreeDecode |
||
131 | add ecx, kLenNumLowSymbols + kLenNumMidSymbols |
||
132 | ret |
||
133 | .0: |
||
134 | mov cl, kLenNumLowBits |
||
135 | shl edx, cl |
||
136 | lea eax, [eax + LenLow*4 + edx*4] |
||
137 | RangeDecoderBitTreeDecode: |
||
138 | ; in: eax->probs,ecx=numLevels |
||
139 | ; out: ecx=length; destroys edx |
||
140 | push edi |
||
141 | xor edx, edx |
||
142 | inc edx |
||
143 | mov edi, edx |
||
144 | xchg eax, edi |
||
145 | @@: |
||
146 | push eax |
||
147 | lea eax, [edi+edx*4] |
||
148 | call RangeDecoderBitDecode |
||
149 | pop eax |
||
150 | adc dl, dl |
||
151 | add al, al |
||
152 | loop @b |
||
153 | sub dl, al |
||
154 | pop edi |
||
155 | mov ecx, edx |
||
156 | ret |
||
157 | |||
158 | ; void __stdcall lzma_decompress( |
||
159 | ; const void* source, |
||
160 | ; void* destination, |
||
161 | ; unsigned dest_length); |
||
162 | lzma_decompress equ _lzma_decompress@12 |
||
163 | public lzma_decompress |
||
164 | lzma_decompress: |
||
165 | push esi edi ebx ebp |
||
166 | mov esi, [esp+4*4+4] ; source |
||
167 | xor ebp, ebp |
||
168 | mov edi, code_ |
||
169 | inc esi |
||
170 | lodsd |
||
171 | bswap eax |
||
172 | stosd |
||
173 | xor eax, eax |
||
174 | dec eax |
||
175 | stosd |
||
176 | stosd |
||
177 | stosd |
||
178 | stosd |
||
179 | xchg eax, esi |
||
180 | stosd |
||
181 | mov ecx, Literal + (LZMA_LIT_SIZE shl (lc+lp)) |
||
182 | mov eax, kBitModelTotal/2 |
||
183 | mov edi, p |
||
184 | rep stosd |
||
185 | mov edi, [esp+4*4+8] ; destination |
||
186 | mov ebx, edi |
||
187 | add ebx, [esp+4*4+12] ; dest_length |
||
188 | .main_loop: |
||
189 | cmp edi, ebx |
||
190 | jae .main_loop_done |
||
191 | mov edx, edi |
||
192 | and edx, posStateMask |
||
193 | push eax ; al = previous byte |
||
194 | mov eax, ebp |
||
195 | shl eax, kNumPosBitsMax+2 |
||
196 | lea eax, [p + IsMatch*4 + eax + edx*4] |
||
197 | call RangeDecoderBitDecode |
||
198 | pop eax |
||
199 | jc .1 |
||
200 | movzx eax, al |
||
201 | if literalPosMask |
||
202 | mov ah, dl |
||
203 | and ah, literalPosMask |
||
204 | end if |
||
205 | shr eax, 8-lc |
||
206 | imul eax, LZMA_LIT_SIZE*4 |
||
207 | add eax, p+Literal*4 |
||
208 | mov cl, 1 |
||
209 | cmp ebp, kNumLitStates |
||
210 | jb .literal |
||
211 | mov dl, [edi + esi] |
||
212 | .lx0: |
||
213 | add dl, dl |
||
214 | setc ch |
||
215 | push eax |
||
216 | lea eax, [eax+ecx*4+0x100*4] |
||
217 | call RangeDecoderBitDecode |
||
218 | pop eax |
||
219 | adc cl, cl |
||
220 | jc .lx1 |
||
221 | xor ch, cl |
||
222 | test ch, 1 |
||
223 | mov ch, 0 |
||
224 | jz .lx0 |
||
225 | .literal: |
||
226 | @@: |
||
227 | push eax |
||
228 | lea eax, [eax+ecx*4] |
||
229 | call RangeDecoderBitDecode |
||
230 | pop eax |
||
231 | adc cl, cl |
||
232 | jnc @b |
||
233 | .lx1: |
||
234 | mov eax, ebp |
||
235 | cmp al, 4 |
||
236 | jb @f |
||
237 | cmp al, 10 |
||
238 | mov al, 3 |
||
239 | jb @f |
||
240 | mov al, 6 |
||
241 | @@: sub ebp, eax |
||
242 | xchg eax, ecx |
||
243 | .stosb_main_loop: |
||
244 | stosb |
||
245 | jmp .main_loop |
||
246 | .1: |
||
247 | lea eax, [p + IsRep*4 + ebp*4] |
||
248 | call RangeDecoderBitDecode |
||
249 | jnc .10 |
||
250 | add eax, (IsRepG0 - IsRep)*4 ;lea eax, [p + IsRepG0*4 + ebp*4] |
||
251 | call RangeDecoderBitDecode |
||
252 | jc .111 |
||
253 | mov eax, ebp |
||
254 | shl eax, kNumPosBitsMax+2 |
||
255 | lea eax, [p + IsRep0Long*4 + eax + edx*4] |
||
256 | call RangeDecoderBitDecode |
||
257 | jc .1101 |
||
258 | cmp ebp, 7 |
||
259 | sbb ebp, ebp |
||
260 | lea ebp, [ebp+ebp+11] |
||
261 | mov al, [edi + esi] |
||
262 | jmp .stosb_main_loop |
||
263 | .111: |
||
264 | add eax, (IsRepG1 - IsRepG0) * 4 ;lea eax, [p + IsRepG1*4 + ebp*4] |
||
265 | call RangeDecoderBitDecode |
||
266 | xchg esi, [rep1] |
||
267 | jnc @f |
||
268 | add eax, (IsRepG2 - IsRepG1) * 4 ;lea eax, [p + IsRepG2*4 + ebp*4] |
||
269 | call RangeDecoderBitDecode |
||
270 | xchg esi, [rep2] |
||
271 | jnc @f |
||
272 | xchg esi, [rep3] |
||
273 | @@: |
||
274 | .1101: |
||
275 | mov eax, p + RepLencoder*4 |
||
276 | call LzmaLenDecode |
||
277 | push 8 |
||
278 | jmp .rmu |
||
279 | .10: |
||
280 | xchg esi, [rep1] |
||
281 | xchg esi, [rep2] |
||
282 | mov [rep3], esi |
||
283 | mov eax, p + Lencoder*4 |
||
284 | call LzmaLenDecode |
||
285 | push kNumLenToPosStates-1 |
||
286 | pop eax |
||
287 | cmp eax, ecx |
||
288 | jb @f |
||
289 | mov eax, ecx |
||
290 | @@: |
||
291 | push ecx |
||
292 | push kNumPosSlotBits |
||
293 | pop ecx |
||
294 | shl eax, cl |
||
295 | shl eax, 2 |
||
296 | add eax, p+PosSlot*4 |
||
297 | call RangeDecoderBitTreeDecode |
||
298 | mov esi, ecx |
||
299 | cmp ecx, kStartPosModelIndex |
||
300 | jb .l6 |
||
301 | push ecx |
||
302 | xor eax, eax |
||
303 | inc eax |
||
304 | shr ecx, 1 |
||
305 | adc al, al |
||
306 | dec ecx |
||
307 | shl eax, cl |
||
308 | mov esi, eax |
||
309 | pop edx |
||
310 | cmp edx, kEndPosModelIndex |
||
311 | jae .l5 |
||
312 | sub eax, edx |
||
313 | shl eax, 2 |
||
314 | add eax, p + (SpecPos - 1)*4 |
||
315 | jmp .l59 |
||
316 | .l5: |
||
317 | sub ecx, kNumAlignBits |
||
318 | ; call RangeDecoderDecodeDirectBits |
||
319 | ;RangeDecoderDecodeDirectBits: |
||
320 | xor eax, eax |
||
321 | .l: |
||
322 | shr dword [range], 1 |
||
323 | add eax, eax |
||
324 | mov edx, [code_] |
||
325 | sub edx, [range] |
||
326 | jb @f |
||
327 | mov [code_], edx |
||
328 | inc eax |
||
329 | @@: |
||
330 | call update_decoder |
||
331 | loop .l |
||
332 | ; ret |
||
333 | mov cl, kNumAlignBits |
||
334 | shl eax, cl |
||
335 | add esi, eax |
||
336 | mov eax, p+Align_*4 |
||
337 | .l59: |
||
338 | ; call RangeDecoderReverseBitTreeDecode_addesi |
||
339 | ;_RangeDecoderReverseBitTreeDecode_addesi: |
||
340 | ; in: eax->probs,ecx=numLevels |
||
341 | ; out: esi+=length; destroys edx |
||
342 | push edi ecx |
||
343 | xor edx, edx |
||
344 | inc edx |
||
345 | xor edi, edi |
||
346 | @@: |
||
347 | push eax |
||
348 | lea eax, [eax+edx*4] |
||
349 | call RangeDecoderBitDecode |
||
350 | lahf |
||
351 | adc edx, edx |
||
352 | sahf |
||
353 | rcr edi, 1 |
||
354 | pop eax |
||
355 | loop @b |
||
356 | pop ecx |
||
357 | rol edi, cl |
||
358 | add esi, edi |
||
359 | pop edi |
||
360 | ; ret |
||
361 | .l6: |
||
362 | pop ecx |
||
363 | not esi |
||
364 | push 7 |
||
365 | .rmu: |
||
366 | cmp ebp, 7 |
||
367 | pop ebp |
||
368 | jb @f |
||
369 | inc ebp |
||
370 | inc ebp |
||
371 | inc ebp |
||
372 | @@: |
||
373 | .repmovsb: |
||
374 | inc ecx |
||
375 | push esi |
||
376 | add esi, edi |
||
377 | rep movsb |
||
378 | lodsb |
||
379 | pop esi |
||
380 | jmp .stosb_main_loop |
||
381 | .main_loop_done: |
||
382 | pop ebp ebx edi esi |
||
383 | ret 12 |
||
384 | |||
385 | section '.bss' data |
||
386 | p rd LZMA_BASE_SIZE + (LZMA_LIT_SIZE shl (lc+lp)) |
||
387 | code_ dd ? |
||
388 | range dd ? |
||
389 | rep1 dd ? |
||
390 | rep2 dd ? |
||
391 | rep3 dd ? |
||
392 | inptr dd ? |
||
393 | previousByte db ? |